generative-ui-specialist

Generative UI Specialist

When to Apply

• Chat interfaces rendering rich components, not just text
• LLM responses with interactive UI (forms, charts, cards, tables)
• Streaming UI that progressively renders during generation
• Agent dashboards with real-time tool visualization

Patterns

• GEN-01 (CRITICAL): Vercel AI SDK 6 -- ToolLoopAgent, createAgentUIStreamResponse, typed message.parts. See rules/vercel-ai-sdk.md
• GEN-02 (CRITICAL): JSON-based rendering -- json-render catalogs, custom registries. See rules/json-render.md
• GEN-03 (HIGH): Streaming chat with tool visualization. See rules/streaming-chat.md
• GEN-04 (HIGH): Structured output -- generateObject/streamObject. See rules/vercel-ai-sdk.md
• GEN-05 (HIGH): Ecosystems -- CopilotKit, assistant-ui, A2UI, Tambo. See rules/ecosystems.md
• GEN-06 (HIGH): Tool approval -- needsApproval patterns. See rules/tool-approval.md

Rules

• GEN-R01 (CRITICAL): Validate LLM JSON before rendering. Use Zod or json-render catalogs. Never dangerouslySetInnerHTML.
• GEN-R02 (CRITICAL): Every registry component needs a fallback for unknown types.
• GEN-R03 (HIGH): Stream everything. Never show blank screens.
• GEN-R04 (HIGH): Show tool status via typed message.parts with loading/complete states.
• GEN-R05 (HIGH): Use stopWhen: stepCountIs(n) to limit agent loops.
• GEN-R06 (MEDIUM): Keep registry small (10-15 components).
• GEN-R07 (MEDIUM): Design for mobile. Test at 375px.
• GEN-R08 (MEDIUM): Use Context7 MCP for current AI SDK docs.
• GEN-R09 (MEDIUM): Use needsApproval instead of custom approval flows.

Anti-Patterns

• Raw JSON dump to users instead of typed components
• No loading/streaming states (blank screens)
• Rendering unsanitized LLM HTML
• Missing error boundaries per rendered block
• Blocking UI with long agent operations
• No conversation persistence across reloads
• Deprecated APIs: maxSteps, parameters, toDataStreamResponse(), ai/rsc
• Untyped toolInvocations instead of typed message.parts